Seque is developed and maintained by DFORUM (Pty) Ltd ("we", "our", "us"), a company registered in South Africa.

Seque exists to stop sensitive information reaching public AI platforms. A tool built for that purpose has to be unusually clear about what it does with the information it sees, so this policy states plainly what is read, what leaves your device, what never does, and what you can ask us to do about it.

It covers the Seque browser extension, the optional local Seque agent, and the Seque control-plane server.

1. The short version

• Your prompts are not sent to us. Classification happens on your device. There is one exception, off by default, described in section 4.
• Flagged values are never transmitted. If Seque detects an identity number, we transmit a one-way fingerprint of it, never the number.
• Audit reporting is opt-in. With consent withheld, records stay on your device.
• Your administrator can see your violations. If your organisation deploys Seque, it can see that you attempted to submit sensitive data, when, and to which platform.
• You can ask us to erase your data, and we erase it automatically 90 days after a subscription ends.

2. Who is responsible for your information

This depends on how you came to use Seque, and it determines who you should approach about your rights.

If your employer deployed Seque, direct access and erasure requests to them first. They hold the relationship with the data and we act on their instruction. We will always help.

3. What Seque reads on your device

On supported AI platforms (ChatGPT, Claude, Gemini, NotebookLM, Microsoft Copilot, Perplexity, Poe, HuggingFace, DeepSeek, Grok, Mistral, Meta AI, Character.AI, You.com, Phind, v0.dev and Bolt.new), the extension inspects:

• Text typed into the prompt field
• Text pasted into it
• Text extracted from files you attach: PDF, Word (DOCX), Excel (XLS/XLSX), and plain-text files including CSV
• AI responses, only if your administrator enables the emotion-inference monitor

This inspection is local. Content is held in memory long enough to classify it and is not written to disk or transmitted. Seque does not read pages other than the supported AI platforms and its own licensing endpoint, and it does not track browsing activity.

What it cannot see. Dictated prompts that a platform submits as audio without rendering text are not inspectable from a browser extension. Images, video and other binary uploads are recorded but their contents are not scanned. These limits are described in the security white paper.

4. What leaves your device

4.1 Licence verification

Organisation identifier, licence signature and validity metadata, so we can confirm the licence is genuine and current. No prompt content is included. This happens at activation, on expiry, and when a licence is replaced, not continuously.

4.2 Audit reporting, only with consent

If, and only if, audit synchronisation is consented to, a record of each policy event is sent to the Seque server. Consent defaults to off: without it, records remain on the device. Each record contains exactly:

• Event type (for example, that a submission was blocked) and severity
• Timestamp and a record identifier
• The hostname of the AI platform involved, for example chatgpt.com, never the page or conversation
• A device identifier
• One-way fingerprints of the flagged values, so repeat occurrences of the same value can be recognised without the value being known
• Classification metadata: which rule matched and what action was taken

The flagged values themselves are removed before transmission. As a second line of defence, the server rejects known content-bearing field names outright and stores only approved fields, so a faulty client cannot cause content to be retained.

4.3 Device check-in

Where a licence is managed centrally, the extension periodically reports a pseudonymised device identifier, the user identifier your organisation issued, licence type, extension version, and whether the extension is correctly installed and covering the sites it should. This is how an administrator can tell that protection is actually running, and how a tampered or removed installation is detected. It contains no prompt content.

4.4 When you remove Seque

Removing the extension sends one final message. Your browser opens a single web address at api.dfum.org carrying your organisation’s identifier, the device identifier, a timestamp and a signature proving the message came from this device. Nothing else is sent, and no prompt content is involved. You will see a page confirming the removal; that page is the whole of what happens next.

This exists so that an administrator can tell the difference between a device that has stopped reporting because Seque was removed and one that has gone quiet for a reason worth investigating: a laptop that was lost, or protection that was disabled without permission. Without it, every removal would look like the second case.

Two things follow that you should know:

• Your administrator is told that you removed it, and when. If your organisation deployed Seque, this is visible to them.
• The web address is recorded in your browser history like any other page you visit, and contains the organisation and device identifiers described above. You can clear it from your history if you prefer.

The signature is only valid for five minutes, and can only be produced by the device it identifies, so it cannot be reused, and it cannot be used to report a removal on somebody else’s behalf.

4.5 The one exception: assisted classification

This applies only to the separately installed Seque agent. The browser extension has no such capability.

An administrator may enable assisted classification, which sends text the local rules could not confidently classify to a third-party AI provider (OpenAI, Anthropic, or an endpoint your organisation nominates) for a second opinion. It is off by default and requires deliberate action at two levels: an administrator enabling it for the organisation, and the agent being configured with the provider details.

When enabled, text is sent only where local rules found little or nothing, which is precisely the text those rules did not flag. Values the local rules did flag are removed first; the remainder is sent in full, truncated to 4,096 characters. Our server does not log or retain it; the provider's own terms then apply.

If assisted classification is not enabled, no prompt content is transmitted anywhere. If you are unsure which applies to you, ask your administrator.

5. What is never transmitted

• Prompt or document content, except through assisted classification (section 4.5)
• Raw flagged values, only one-way fingerprints
• Page contents, URLs beyond the hostname, or conversation history
• Browsing activity outside the supported AI platforms
• Passwords. Administrator sign-in uses your organisation's own identity provider, so we hold no password to lose

We do not sell personal information, use it for advertising, or use your data or audit records to train machine-learning models.

6. Where information is stored, and for how long

On your device

Classifier rules, licence information, settings and the local audit log are held in browser extension storage, capped at 2,000 entries. Organisation-wide policy arrives through managed storage in enterprise deployments. Sensitive stored values are encrypted (AES-256-GCM).

An honest limit: because the extension must decrypt unattended when your browser starts, the key material lives alongside the data it protects. This prevents values being lifted to another machine and stops casual inspection, but it is not a defence against someone with full access to an unlocked profile. Full-disk encryption on your device is the control that answers that risk.

On our servers

Erasure removes the data, not the fact that events occurred. Audit entries are linked by a hash chain that makes tampering detectable, and that chain is shared across records. Deleting rows outright would break it for customers who are not being erased. Entries are therefore emptied rather than removed: they retain no identifier, hostname or fingerprint that can be traced to you, and the process cannot be reversed.

7. Who else is involved

A current list of sub-processors and a Data Processing Agreement are available on request.

8. Where your information is processed

Seque supports regional deployment, so account, licensing and audit data can be held within a chosen jurisdiction, including South Africa, for POPIA. Devices are directed to their regional server by their licence, and that address is checked against a permitted list, so a licence cannot redirect data to an arbitrary destination.

Some services sit outside the deployment and may involve a cross-border transfer: email delivery, bot protection on the trial form, error reporting, and assisted classification where enabled. Where your organisation requires data to stay in a jurisdiction, these are configurable, and assisted classification should be assessed against POPIA section 72 or GDPR Articles 44–49 before it is turned on.

9. Your rights

Under POPIA and, where it applies, the GDPR, you may:

• Ask what we hold about you and receive a copy
• Ask us to correct anything inaccurate
• Ask us to erase your information (POPIA section 24; GDPR Article 17). Seque includes a built-in erasure function for exactly this
• Object to processing, or ask that it be restricted
• Receive your data in a portable form, audit records export as CSV
• Complain to a regulator: the Information Regulator (South Africa), or your local supervisory authority under the GDPR

Where your employer deployed Seque, ask them first, we act on their instructions. Contact us at connect@dfum.org and we will respond within 30 days.

Our lawful basis for processing is the legitimate interest of an organisation in preventing loss of confidential and personal data, and the performance of our contract with a customer. Audit reporting to our servers additionally rests on consent, which is why it is off until granted and can be withdrawn.

10. Sensitive categories of information

Seque can be configured to detect special personal information under POPIA section 26 and special-category data under GDPR Article 9, including health information, medical scheme details, disclosures relating to mental health, and neural or cognitive data.

These classifiers exist to prevent such information being submitted to a public AI platform. Detection happens on your device, and where a match is reported the record contains the category matched and a one-way fingerprint, never the disclosure itself. The optional emotion-inference monitor works the same way: it warns that an AI response appears to be inferring emotional or psychological state, and records that it did so.

Organisations enabling these classifiers should confirm they have a lawful basis for the processing, since the detection itself involves the category concerned.

11. If your organisation deployed Seque

You should know what your administrator can see. Where audit reporting is enabled, they can see that you attempted to submit something matching a policy rule, what type it was, when, and to which AI platform. They cannot see the content of your prompt or the flagged value itself.

Administrators can also deploy organisation-wide rules that you cannot see or disable, require blocking rather than warning, and enforce these settings centrally. Seque does not override your organisation's own privacy obligations, and its use as a monitoring tool remains subject to the employment and privacy law of your jurisdiction.

12. Security

• Content is classified locally rather than uploaded for inspection
• Flagged values are replaced with one-way fingerprints before any transmission
• Licences are cryptographically signed and verified; an invalid signature is refused
• Requests to our servers are individually signed, time-limited and single-use, so an intercepted request cannot be replayed
• Sensitive database fields are encrypted, and audit records are hash-chained so tampering is detectable
• Administrator access uses your organisation's identity provider, with role-based permissions and optional multi-factor authentication

Our security design, including its limitations, is described in the Seque security white paper. If you believe you have found a vulnerability, please write to connect@dfum.org.

13. Children

Seque is a workplace tool, intended for organisational and professional use. It is not designed or marketed for use by children, and we do not knowingly collect information from anyone under 18.

14. Changes to this policy

We may update this policy. Where a change materially affects how your information is handled, we will notify customers by email or an in-product notice at least 14 days before it takes effect. The current version is always published with the product and on our website.

15. Contact

DFORUM (Pty) Ltd (Data for Optimised Reasoning and Unified Modelling)

Privacy, access and erasure requests: connect@dfum.org

Website: dfum.org

South African users may also complain to the Information Regulator (South Africa) at inforegulator.org.za.